Classification of Zambian grasslands using random forest feature importance selection during the optimal phenological period

• Published in: Ecological indicators Vol. 135; p. 108529
• Main Authors: Zhao, Yifan, Zhu, Weiwei, Wei, Panpan, Fang, Peng, Zhang, Xiwang, Yan, Nana, Liu, Wenjun, Zhao, Hao, Wu, Qirui
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2022; Elsevier
• Subjects: algorithms; biomass; carbon sequestration; Central Africa; data collection; GEE; grasses; Grassland classification; grasslands; Internet; land cover; Optimal feature selection; phenology; PROBA-V; range management; savannas; time series analysis; vegetation index; Zambia; Central Africa; Zambia; Optimal feature selection; Grassland classification; PROBA-V; GEE
• DOI: 10.1016/j.ecolind.2021.108529
• ISSN: 1470-160X

•The cloud computing capability of GEE improves the efficiency of obtaining national-scale grassland information.•Time series NDVI can assist in the selection of optimal phenological periods and improve computational efficiency.•Optimal feature selection can reduce the number of features and improve the classification accuracy.•Elevation was found to be the most critical feature for the classification of Zambian grasslands. It is important to conduct grassland resource surveys for the scientific management of grassland resources. Currently, remote sensing technology is widely used to classify land cover. The fine classification datasets of grasslands with high spatial and temporal resolutions are very necessary for scientific research. In order to use remote sensing data conveniently, this study selected the Google Earth Engine platform to select 100-m resolution PROBA-V remote sensing images from 2018 of Zambia, in central Africa. The differences in the normalized vegetation index time-series curves of the different types of grasslands were combined, and June to October was identified as the best phenological classification period. Using the random forest feature importance selection algorithm, the original feature indices and identification of the different grass types were optimized. The results indicate that using the optimal feature combination selected by the random forest feature importance selection algorithm to refine the classification of grasslands improves computational efficiency with an overall accuracy of 83%, which is 3% higher than that of the original feature combination. Among the optimal feature combinations, elevation contributes the most to the improvement classification accuracy. The most significant improvement in the producer’s accuracy was found for grassland (30% increase) and savanna (22% increase). Adjustment of the appropriate phenological periods according to the seasonal characteristics of different regions, the methodology established in this study can be easily applied to other areas for the fine classification of grasslands and the subsequent calculation of grassland biomass and carbon storage.